fengyanzi/SHD25

SHD-25

SHD-25 is a hyperspectral ship detection dataset collected from EO-1 Hyperion for small vessels in coastal and harbor scenes.

• Paper: https://www.sciencedirect.com/science/article/pii/S0031320326009210
• DOI: https://doi.org/10.1016/j.patcog.2026.113956
• Dataset: https://huggingface.co/datasets/fengyanzi/SHD25
• Code: https://github.com/nal793/DCH-Net

The release contains 120 hyperspectral chips of (256\times256) pixels with 242 bands, YOLO boxes for the ship class, an RGB quicklook for conventional detectors, and a preview with the ground-truth boxes drawn in yellow.

What is in the set

SHD-25 supports ship detection in complex maritime environments. The imagery comes from the Hyperion sensor on EO-1 and covers coasts and harbors in China, Singapore, Panama, South Korea, Nigeria, the Netherlands, and the United States. There are 120 chips. Most ships are small, with a spatial extent on the order of 10 to 100 pixels.

Each chip is (256\times256). The cube stores the 242 Hyperion bands after radiometric calibration. DCH-Net uses the subset

[8–56, 81–118, 132–162, 187–213]

which holds 145 bands and leaves out wavelengths that are heavily affected by atmospheric absorption. The chips do not carry a map projection.

Yellow boxes in the figures below are the ground truth, drawn on the RGB quicklook for inspection.

Shanghai shanghai_01	Panama panama_01

Los Angeles losangeles_01	Rotterdam rotterdam_01

Layout

Images and labels share one file stem, in the same spirit as a YOLO dataset. Hyperspectral cubes, RGB quicklooks, and previews sit beside the labels.

SHD-25/
├── README.md
├── LICENSE
├── shd25.yaml
├── classes.txt                 # ship
├── hsi/<file_name>.tif         # hyperspectral cube
├── rgb/<file_name>.png         # RGB quicklook
├── labels/<file_name>.txt      # YOLO boxes
├── previews/<file_name>.jpg    # quicklook with yellow boxes
└── meta/
    ├── index.csv
    ├── stats.json
    └── checksums.sha256

A stem looks like shanghai_01 or panama_03. Each of the four data folders has 120 files. sample_id runs from SHD25_001 to SHD25_120 in meta/index.csv.

Path	Count	Format	Role
hsi/	120	TIFF, uint16, 256×256×242, uncompressed	hyperspectral detectors
rgb/	120	PNG, 8-bit RGB, 256×256	conventional detectors
labels/	120	one object per line	training and evaluation
previews/	120	JPEG with yellow boxes	visual check
Total	120 chips / 786 boxes	cubes about 3.6 GB, the rest about 10 MB

shd25.yaml points a YOLO loader at rgb/ with the class name ship.

Index

meta/index.csv has one row per chip.

Field	Meaning
sample_id	SHD25_001 through SHD25_120
file_name	stem shared by the cube, quicklook, label, and preview
scene_id	scene name and filename prefix
scene_index	index inside that scene
place_en / place_zh	port or city
country_en / country_zh	country or region
original_name	name used while the chip was collected
n_objects	number of ship boxes

Scenes

Prefix	Place	Region	Images	Boxes
shanghai	Shanghai	China	23	264
hongkong	Hong Kong	China	16	58
singapore	Singapore	Singapore	32	178
singapore2	Singapore, second scene	Singapore	23	194
panama	Panama	Panama	11	26
busan	Busan	South Korea	4	9
lagos	Lagos	Nigeria	1	4
rotterdam	Rotterdam	the Netherlands	4	26
losangeles	Los Angeles	United States	4	23
ship5	ship5	—	2	4
Total			120	786

Every chip has at least one box. The busiest chip has 43. The mean is 6.55 boxes per chip.

Hyperspectral cubes

• Shape (256, 256, 242), axis order height, width, band.
• Dtype uint16, uncompressed, about 30.3 MB per chip.
• Band order follows the 242 Hyperion bands.
• The 145 bands used by DCH-Net are the 1-based IDs below. Subtract 1 to index the cube.

Band IDs (inclusive)	0-based indices
8–56	7–55
81–118	80–117
132–162	131–161
187–213	186–212

from shd25.bands import paper_band_indices

cube_sel = cube[:, :, paper_band_indices()]  # (256, 256, 145)

Labels

classes.txt contains a single name, ship. Label files use the same stem as the image. Each line is one object:

class cx cy w h

class is 0. cx and cy are the box center. w and h are the width and height. All four values are divided by 256 and lie in ([0, 1]).

The start of labels/shanghai_01.txt:

0 0.375 0.859376 0.046876 0.039062
0 0.482422 0.642578 0.035156 0.035156
0 0.048828 0.408204 0.050782 0.042968

The yellow rectangles in previews/ are these boxes drawn on the quicklook. Use labels/ for training and for computing metrics.

RGB quicklooks

rgb/ holds 8-bit PNG images for detectors that take a three-channel input. Pair each file with the label of the same stem. The quantitative spectra are in hsi/.

Loading

Pass this directory to the reader after installing the DCH-Net package:

from shd25 import SHD25, select_paper_bands

dataset = SHD25("SHD25")
sample = dataset[0]
print(sample["id"], sample["hsi"].shape, select_paper_bands(sample["hsi"]).shape)
print(sample["boxes"][:2])

Cubes load with tifffile as uint16 arrays of shape (256, 256, 242).

meta/checksums.sha256 follows the GNU sha256sum layout. Paths are relative to this directory.

License and citation

The dataset is released under CC BY 4.0. EO-1 Hyperion imagery is a public satellite product. The chips and the ship boxes are provided by the authors of DCH-Net. Please cite:

@article{niu2026dchnet,
  title   = {DCH-Net: A hyperspectral object detection network with differential convolution and spectral gradient fusion},
  author  = {Niu, Ailin and Yan, Xinyu and Wang, Qiang and Chen, Jiuchen and Han, Xiaolin and Xu, Qizhi},
  journal = {Pattern Recognition},
  volume  = {180},
  pages   = {113956},
  year    = {2026},
  doi     = {10.1016/j.patcog.2026.113956}
}

Authors: Ailin Niu, Xinyu Yan, Qiang Wang (corresponding author), Jiuchen Chen, Xiaolin Han, and Qizhi Xu, School of Mechatronics Engineering, Beijing Institute of Technology. This work was supported by the National Natural Science Foundation of China under Grant 62572044.